Abstract / Summary
The Centiloid (CL) scale has become a crucial quantitative metric for β-amyloid (Aβ) PET in Alzheimer's disease. However, variability in processing methods complicates its clinical adoption. This study systematically evaluated the reliability and clinical utility of different CL processing methods to support their application in clinical practice. 18 F-florbetaben (FBB) and 18 F-florbetapir (FBP) datasets from the Global Alzheimer's Association Interactive Network (GAAIN) and 127 subjects (64 with 18 F-FBB and 63 with 18 F-FBP) from a local database were analysed. CL values were derived from three software platforms, SPM, PMOD, and BTXBrain, employing distinct CL processing methods. First, the robustness of the processing methods was evaluated by comparing CL values derived from public database images against published reference values using linear regression. Subsequently, CL values of local datasets were generated, and the amyloid deposition status was determined with cut-off values of 12 or 24. Diagnostic performance was evaluated using visual assessment results as the reference standard, with sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy calculated accordingly. MRI-based methods consistently met the established criteria (slope: 0.98–1.02; intercept: − 2–2 CL; R 2 > 0.98) and exhibited notable diagnostic performance at CL cut-off values of 12 and 24 (sensitivity/NPV = 100%; accuracy ≥ 94%). PET-only processing methods were highly tracer-specific. For FBB, SPM and BTXBrain exhibited robust and comparable outcomes, with all key metrics exceeding 90%. For FBP, BTXBrain demonstrated superior diagnostic performance over SPM and PMOD in terms of accuracy (82.5–85.7%), sensitivity (77.3–84.1%), and NPV (64.3–70.8%). SPM showed intermediate performance in terms of accuracy (73.0–81.0%), sensitivity (63.6–77.3%) and NPV (52.9–63.0%). Although PMOD achieved perfect specificity and PPV (100%), it exhibited the lowest accuracy (60.3–66.7%), sensitivity (43.2–52.3%) and NPV (43.2–47.5%). The findings indicate that MRI-based methods provide highly consistent and robust diagnostic performance, whereas BTXBrain has a relative advantage in PET-only processing, particularly with FBP tracers, although further multicentre validation is warranted.